Gemma 4 models
44 Gemma 4 models from Google on Hugging Face, from 420M to 32.7B parameters, published by Google in BF16, with contexts from 128K to 256K tokens. The smallest official model, gemma-4-26B-A4B-it-assistant, needs about 0.9 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Part of the Gemma series ยท Previous generation: Gemma 3n
Pick a size
One row per official Gemma 4 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| gemma-4-26B-A4B-it-assistant | 420M | 0.9 GB | 0.5 GB | 0.2 GB | RTX 3070 | $0.088/hr | 0.27 GB |
| gemma-4-31B-it-assistant | 470M | 1.0 GB | 0.5 GB | 0.3 GB | RTX 3070 | $0.088/hr | 0.55 GB |
| gemma-4-E2B-it | 5.1B | 11.5 GB | 5.7 GB | 2.9 GB | RTX A4000 | $0.137/hr | 0.23 GB |
| gemma-4-E4B-it | 8.0B | 17.9 GB | 8.9 GB | 4.5 GB | RTX A5000 | $0.176/hr | 0.47 GB |
| gemma-4-12B-it | 12.0B | 26.7 GB | 13.4 GB | 6.7 GB | RTX A6000 | $0.363/hr | 2.3 GB |
| gemma-4-26B-A4B-it | 25.8B | 57.7 GB | 28.8 GB | 14.4 GB | A100 | $1.31/hr | 1.4 GB |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | 26.5B | 59.3 GB | 29.7 GB | 14.8 GB | A100 | $1.31/hr | 1.4 GB |
| gemma-4-31B-it | 31.3B | 69.9 GB | 35.0 GB | 17.5 GB | A100 | $1.31/hr | 5.8 GB |
| gemma-4-31B | 32.7B | 73.1 GB | 36.5 GB | 18.3 GB | A100 | $1.31/hr | 5.8 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (15)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| gemma-4-26B-A4B-it-assistant | 420M | BF16 | 0.9 GB | RTX 3070 | $0.088/hr |
| gemma-4-12B-it-assistant | 423M | BF16 | 0.9 GB | RTX 3070 | $0.088/hr |
| gemma-4-12B-it-qat-q4_0-unquantized-assistant | 423M | BF16 | 0.9 GB | RTX 3070 | $0.088/hr |
| gemma-4-31B-it-assistant | 470M | BF16 | 1.0 GB | RTX 3070 | $0.088/hr |
| gemma-4-31B-it-qat-q4_0-unquantized-assistant | 470M | BF16 | 1.0 GB | RTX 3070 | $0.088/hr |
| gemma-4-E2B-it | 5.1B | BF16 | 11.5 GB | RTX A4000 | $0.137/hr |
| gemma-4-E4B | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| gemma-4-E4B-it | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| gemma-4-12B | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| gemma-4-12B-it | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| gemma-4-12B-it-qat-q4_0-unquantized | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| gemma-4-26B-A4B-it | 25.8B | BF16 | 57.7 GB | A100 | $1.31/hr |
| gemma-4-26B-A4B-it-qat-q4_0-unquantized | 26.5B | BF16 | 59.3 GB | A100 | $1.31/hr |
| gemma-4-31B-it | 31.3B | BF16 | 69.9 GB | A100 | $1.31/hr |
| gemma-4-31B | 32.7B | BF16 | 73.1 GB | A100 | $1.31/hr |
Quantized versions (13)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| gemma-4-E4B-it-W4A16-AutoRound-GPTQ | 4.0B | BF16 | 8.9 GB | RTX A4000 | $0.137/hr |
| gemma-4-E2B-it-unsloth-bnb-4bit | 5.1B | BF16 | 11.5 GB | RTX A4000 | $0.137/hr |
| gemma-4-E2B-it-qat-w4a16-ct | 5.6B | BF16 | 12.4 GB | RTX A4000 | $0.137/hr |
| Gemma-4-E4B-DECKARD-HERETIC-NVFP4 | 6.2B | BF16 | 13.9 GB | RTX A4000 | $0.137/hr |
| gemma-4-E4B-it-unsloth-bnb-4bit | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| gemma-4-E4B-it-AWQ-INT4 | 8.2B | F16 | 18.3 GB | RTX A5000 | $0.176/hr |
| gemma-4-E4B-it-qat-w4a16-ct | 8.7B | BF16 | 19.5 GB | RTX A5000 | $0.176/hr |
| gemma-4-12B-it-FP8-Dynamic | 13.0B | F8_E4M3 | 14.5 GB | RTX 5060 Ti | $0.180/hr |
| Gemma-4-31B-IT-NVFP4 | 20.9B | BF16 | 46.6 GB | RTX A6000 | $0.363/hr |
| rune-26b-a4b-GGUF | 25.8B | BF16 | 57.7 GB | A100 | $1.31/hr |
| gemma-4-26B-A4B-it-FP8-dynamic | 26.5B | F8_E4M3 | 29.7 GB | RTX 5000 Ada | $0.539/hr |
| gemma-4-31B-it-FP8-block | 31.3B | F8_E4M3 | 35.0 GB | L40 | $0.759/hr |
| gemma-4-31B-it-FP8-dynamic | 31.3B | F8_E4M3 | 35.0 GB | L40 | $0.759/hr |
Fine-tunes and community models (16)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| gemma-4-26B-A4B-it-DFlash | 430M | BF16 | 1.0 GB | RTX 3070 | $0.088/hr |
| gemma-4-31B-it-DFlash | 1.5B | BF16 | 3.4 GB | RTX 3070 | $0.088/hr |
| gemma4-e4b-claims-comparison | 7.9B | BF16 | 17.7 GB | RTX A5000 | $0.176/hr |
| gemma-4-E4B-it-OBLITERATED | 8.0B | F16 | 17.9 GB | RTX A5000 | $0.176/hr |
| gemma-4-E4B-it-uncensored-heretic | 8.0B | BF16 | 17.9 GB | RTX A5000 | $0.176/hr |
| Jev-Omni | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| humanizer | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| Gemma-4-12B-OBLITERATED | 12.0B | BF16 | 26.7 GB | RTX A6000 | $0.363/hr |
| gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 | 12.0B | F16 | 26.7 GB | V100 | $0.220/hr |
| GEV-26B-Decide | 25.8B | BF16 | 57.7 GB | A100 | $1.31/hr |
| Orion-26B-A4B-v1 | 25.8B | BF16 | 57.7 GB | A100 | $1.31/hr |
| Orion-26B-A4B-v1.1 | 25.8B | BF16 | 57.7 GB | A100 | $1.31/hr |
| SOV-Transcribe-1 | 26.6B | BF16 | 59.5 GB | A100 | $1.31/hr |
| Artemis-31B-v1.2 | 31.3B | BF16 | 69.9 GB | A100 | $1.31/hr |
| OTel-2.0-LLM-31B-IT | 32.1B | BF16 | 71.8 GB | A100 | $1.31/hr |
| gemma-4-31B-it-uncensored | 32.7B | BF16 | 73.1 GB | A100 | $1.31/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune Gemma 4
- How to run Gemma 4: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Gemma 4: LoRA and QLoRA memory per model, and the license limits on derivatives.